A module for compressing a speech signal is called an encoder, and a module for decompressing a compressed speech signal is called a decoder. The most basic speech codec is an ITU-T G.711 codec which samples an input signal at 8 kHz and quantizes the sampled input signal to 8 bits. Where, in order to increase quantization efficiency, an A-law log quantizer as shown in Equation 1 or a u-law log quantizer as shown in Equation 2 is used.
                                                                        AL                ⁡                                  (                  x                  )                                            =                              Ax                                  1                  +                                                            log                      10                                        ⁡                                          (                      A                      )                                                                                                                                              for                ⁢                                                                  ⁢                0                            ≤              x              ≤                              1                A                                                                                                        AL                ⁡                                  (                  x                  )                                            =                                                1                  +                                                            log                      10                                        ⁡                                          (                      Ax                      )                                                                                        1                  +                                                            log                      10                                        ⁡                                          (                      A                      )                                                                                                                                              for                ⁢                                                                  ⁢                                  1                  A                                            ≤              x              ≤              1                                                          (        1        )                                          uL          ⁡                      (            x            )                          =                              sign            ⁡                          (              x              )                                ⁢                                                    V                0                            ⁢                                                log                  10                                ⁡                                  [                                      1                    +                                                                  u                        ⁢                                                                            x                                                                                                                      V                        0                                                                              ]                                                                                    log                10                            ⁡                              [                                  1                  +                  u                                ]                                                                        (        2        )            
A conventional log quantizer as described above applies different quantizing intervals according to the magnitudes of input signals. For example, a relatively wide quantizing interval is set for a signal having a small magnitude, and a relatively narrow quantizing interval is set for a signal having a large magnitude, that is, a signal highly likely to be generated. Accordingly, the efficiency of quantization is increased.
It is well known that quantization noise is evenly distributed over the entire bandwidth. However, according to the characteristics of human hearing, a quantization error existing in a segment of a signal having a large magnitude is not clearly heard as it is buried in the signal, and a quantization error existing in a segment of a signal having a small magnitude is easily heard as a noise. Accordingly, not only a speech segment but also a silence segment needs to be effectively coded because the coding of the silence segment affects the overall performance of a codec. In other words, noise caused by a quantization error in a silence segment may affect the overall sound quality.
A codec may have different performances according to the magnitudes of an input signal. In order to evaluate the performance of a speech codec, signals having different magnitudes, for example, signals of −16, −26, and −36 dBoV, are usually evaluated. In other words, a codec evaluates how its performance varies according to a change in the amplitude of an input signal.
In a codec such as G.711 or G.722, noise is generated due to a quantization error with respect to an input signal of −36 dBoV. In particular, a quantization error generated in a silence segment of the input signal serves as a factor in reducing the overall quality of the codec. Results of a subjective hearing test show that a mean opinion score (MOS) with respect to the input signal of −26 dBoV is higher than −36 dBoV.